You Can’t Sell an Agent
A few weeks ago I wrote that AI productivity is real but the scaling isn’t. Teams get 2x, 3x, sometimes 10x on specific tasks. The wins are legible, measurable, reproducible. But they don’t aggregate into business outcomes. The P&L barely moves.
This piece is about why.
The 10% problem
When someone says “we built an agent,” what they usually built is the agent — the LLM, plus a harness to run it, plus memory and accumulated experience. That’s maybe 10% of what needs to exist for the deployment to actually work.
The harness is the technical layer developers build to make a model behave like a system. Code that loads context, hands it tools, catches its errors, logs what it did. Open-strix is one example, open-sourced — the harness running the agents I work with, including the one ghostwriting this post.
Past the agent is the organizational plumbing. Reporting paths. Escalation routes. Exception handling. Accountability lines. Handoff protocols. The operational tissue that decides what the agent does when it doesn’t know what to do, who gets the output, who owns the failures, who corrects the drift, and who explains any of it to an auditor six months later.
None of that ships with the agent. And none of it is what the AI vendors are selling.
Copilot’s fifteen percent
Microsoft has deployed Copilot into a large share of the Fortune 500. Surveys consistently show usage rates stuck around 10-15% of seats, two years in. Not because the tool is bad — it’s excellent for individual tasks. Because seats don’t map to workflows, and workflows are what actually need to change.
This isn’t a Microsoft problem. It’s structural. Every enterprise AI pilot stalls at the same place: the agent does the thing, but the organization around it can’t absorb what the agent produces. It can’t route exceptions back to humans fast enough. It can’t explain its outputs to auditors. It can’t decide what happens when the agent disagrees with a specialist. It can’t cut a check to someone whose job it is to watch the agent at 3am.
The pilots don’t fail from capability gaps. They fail from regulatory mismatch — the organization has no tissue to absorb the new component.
What the product actually is
If the agent is 10%, the product isn’t the agent. The product is the re-architected workflow. The redesigned handoffs. The new reporting cadence. The training that turns a specialist into someone who supervises an agent instead of doing the task directly. The measurement infrastructure that tells you when the agent is drifting, under-firing, or quietly replaced by a Slack group chat.
This doesn’t SKU cleanly. You can’t put “workflow redesign for legal claims triage, mid-market carrier, Q3 implementation” on a pricing page. There’s no self-serve path. There’s no bundle. Every deployment is bespoke at the organizational layer, even when the underlying agent is off-the-shelf.
The industry has a name for businesses that deliver this kind of outcome: firms that understand the work.
The consulting-shaped answer
The winning business model at the application layer isn’t SaaS. It’s consulting-shaped — but not in the legacy sense of twelve-month engagements and thousand-page decks. In a new sense: diagnose, implement, enable. Find where the organization actually gets friction from. Build the agent that targets it. Rebuild the workflow around the agent-augmented team. Train the operators. Instrument the outcome. Hand it off and leave.
The capital structure of a SaaS company — heavy product, low service — doesn’t fit this work. The margins are different. The repeatability is different. But the value is there, and it’s capturable by firms structured for it.
This is partly why the application layer hasn’t crowned a winner yet. The people building SaaS can’t get to the work. The people doing the work look like services businesses, which VCs have been trained to discount.
Why the platforms can’t win here
OpenAI, Anthropic, Google all have application-layer ambitions. But the thing that would make them win the application layer is also what they’re structurally worst at: organizational fit.
Tokens aren’t the bottleneck. Reasoning quality isn’t the bottleneck. The bottleneck is whether an underwriting team can absorb a claims-triage agent into their existing specialist hierarchy without blowing up the cycle time on complex cases. That’s not a model problem. It’s a process problem, and it only shows up when you’re inside the organization, talking to the person whose comp is tied to that cycle time.
Platform companies are optimized for scale. Organizational fit is intrinsically non-scalable. You can’t ship it in a model release.
The frame shift
“Buy an AI agent” is the wrong frame for enterprise buyers. The right frame is: rebuild this workflow around an AI-augmented team. The agent is a tool inside that rebuild. The rebuild is the product.
This matters for how you sell. If you’re pitching an agent, you’re pitching the 10%. The buyer has to do the other 90% themselves, and they don’t know how, and the pilot stalls. If you’re pitching the workflow redesign with the agent as an included component, you’re pitching what the buyer actually needs.
The math changes. So does the sales motion. So does who you hire.
Who does this work
Someone has to actually do it — diagnose the organization, shape the agent against it, land the rebuild. That someone is an operator, and the role is narrower than most people think.
An operator isn’t a generalist who’s read about AI. It’s someone grounded in two places at once:
Agents. Not just how to prompt them — how they fail, how harnesses shape their behavior, what memory does and doesn’t do, where the capability curve is bending, and where it isn’t. The actual texture of the thing.
The subject matter. The real work — legal ops, marketing, underwriting, claims, whatever the organization does to make money. Not as a consultant who can google the domain, but as someone who’s been inside the work long enough to see where it actually breaks, and why.
The operator’s job is to deeply understand the organizational context, then frame an agent against it in a way that papers over the ugliness and solves a real problem. And — maybe more importantly — to know which problems are worth solving. Most aren’t.
That’s a rare combination. The AI people don’t know the work. The domain people don’t know the agents. The generalists know neither. Most organizations are full of people who know one side and assume the other is easy. It isn’t.
If you’re buying, this is the person you’re looking for. If you’re hiring, this is the person you’re looking for. If you’re investing, the bet is on the people who ARE this person — or who know how to build teams of them.
Where do you find such a person?
The productive version of the boring answer
AI deployments fail from organizational mismatch, not capability. That sounds like a boring old observation wearing new clothes. It isn’t quite. The observation is old. What’s new is that the platform companies have pushed capability so far forward that organizational fit is now the only remaining bottleneck at the application layer. Everything else has been solved or commoditized.
Which means the winners aren’t going to be the platforms. They’re going to be the operators who know how to land agents inside organizations that weren’t built for them.
You can’t sell an agent. You can sell the workflow that makes the agent worth deploying. That’s the product.